1. Trang chủ
  2. » Giáo án - Bài giảng

averaging not internal noise limits the development of coherent motion processing

13 4 0

Đang tải... (xem toàn văn)

Tài liệu hạn chế xem trước, để xem đầy đủ mời bạn chọn Tải xuống

THÔNG TIN TÀI LIỆU

Thông tin cơ bản

Định dạng
Số trang 13
Dung lượng 2,02 MB

Các công cụ chuyển đổi và chỉnh sửa cho tài liệu này

Nội dung

The traditional motion coherence paradigm cannot distinguishbetweenlocalandgloballimitstomotion per-ception and has hence obscured our understanding of whatlimitsglobalmotionprocessingdu

Trang 1

jo u rn al h om ep age : ht t p : / / w w w e l s e v i e r c o m / l o c a t e / d c n

Catherine Manninga,∗, Steven C Dakinb,c, Marc S Tibberb, Elizabeth Pellicanoa

a Centre for Research in Autism and Education (CRAE), Institute of Education, University of London, 55-59 Gordon Square,

Institute of Education, London WC1H 0NU, UK

b UCL Institute of Ophthalmology, University College London, Bath Street, London EC 1V9, UK

c NIHR Biomedical Research Centre at Moorfields Eye Hospital, 162 City Road, London EC 1V 2PD, UK

a r t i c l e i n f o

Article history:

Received 5 February 2014

Received in revised form 16 July 2014

Accepted 18 July 2014

Available online 1 August 2014

Keywords:

Visual development

Motion processing

Direction discrimination

a b s t r a c t

Thedevelopmentofmotionprocessingisacriticalpartofvisualdevelopment,allowing childrentointeractwithmovingobjectsandnavigatewithinadynamicenvironment How-ever,globalmotionprocessing,whichrequirespoolingmotioninformationacrossspace, developslate,reachingadult-likelevelsonlybymid-to-latechildhood.Thereasons under-lyingthisprotracteddevelopmentarenotyetfullyunderstood.Inthisstudy,wesoughtto determinewhetherthedevelopmentofmotioncoherencesensitivityislimitedbyinternal noise(i.e.,imprecisioninestimatingthedirectionsofindividualelements)and/orglobal poolingacrosslocalestimates.Tothisend,wepresentedequivalentnoisedirection dis-criminationtasksandmotioncoherencetasksatbothslow(1.5◦/s)andfast(6◦/s)speeds

tochildrenaged5,7,9and11years,andadults.Weshowthat,aschildrengetolder,their levelsofinternalnoisereduce,andtheyareabletoaverageacrossmorelocalmotion esti-mates.Regressionanalysesindicated,however,thatage-relatedimprovementsincoherent motionperceptionaredrivensolelybyimprovementsinaveragingandnotbyreductions

ininternalnoise.Ourresultssuggestthatthedevelopmentofcoherentmotionsensitivity

isprimarilylimitedbydevelopmentalchangeswithinbrainregionsinvolvedinintegrating motionsignals(e.g.,MT/V5)

©2014TheAuthors.PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCC

BYlicense(http://creativecommons.org/licenses/by/3.0/)

The processing of motion is a critical partof visual

development,allowingchildren totrackmoving objects

withtheireyes, toreachfor and graspobjects thatare

in motion, and to navigate within a dynamic world

Motionprocessing contributestoa range ofelementary

visualfunctionsincludingthesegmentationofscenesinto

differentobjects and surfaces, the perception of depth,

∗ Corresponding author Tel.: +44 207 331 5135.

E-mail address: c.manning@ioe.ac.uk (C Manning).

the registration of trajectories and the identification of objects.Often,it isimportant tocombinemotion infor-mationacrossspace,for exampleinorder todetermine theoveralldirectionofaflockofbirds,eachofwhichwill

befollowinga differentmotiontrajectory.Thisability– termed global motion processing – is typically tested experimentally using the motion coherence paradigm (Newsome and Paré, 1988), which requires observers

tojudgethedirectionof coherentlymoving dotsinthe presenceofrandomlymovingnoisedots

Given theimportanceof motionprocessing invisual development,itisperhapsunsurprisingthatsomeaspects

ofmotionprocessing(e.g.,directionalselectivity)develop

http://dx.doi.org/10.1016/j.dcn.2014.07.004

1878-9293/© 2014 The Authors Published by Elsevier Ltd This is an open access article under the CC BY license ( http://creativecommons.org/

Trang 2

early in life (Wattam-Bell, 1991, 1992; see Braddick

et al.,2003,for review).However,othertypes ofvisual

motionprocessingfollowaprotracteddevelopmentand

only reach adult-like levels by mid-to-late childhood

For example, the minimum speed required to support

perception of motion-defined form and the maximum

displacementsupportingperceptionofmovementmature

byaround7–8years(Haywardetal.,2011;Parrishetal.,

2005), motion coherence thresholds reach adult-like

levelsbetween10and14years(Gunnetal.,2002;Hadad

etal.,2011)andspeeddiscriminationabilitiesarenotyet

fullyadult-likeby11years(Manningetal.,2012).Such

motionprocessing abilities rely primarilyonthe dorsal

pathway (Milner and Goodale, 1995), which originates

frommotion-sensitiveneuronsinareaV1,andprojectsto

extrastriateareasincludingMT/V5.WhileneuronsinV1

cansignalthepresenceoflocalmotion(HubelandWiesel,

1962), neuronsin V5 play a key role in global motion

processing,astheyhavelargerreceptivefieldscapableof

integratinginputsfromV1(Mikamietal.,1986)

Adultstudiesofvisualmotionprocessingsuggestthe

existence of atleast two distinct systems tunedto

dif-ferentrangesofspeed(Burretal.,1998;Edwardsetal.,

1998;Thompsonetal.,2006;alsoseereviewbyBurrand

Thompson, 2011), which may follow different

develop-mentaltrajectoriesinthematuringbrain.Haywardetal

(2011)reportedgreaterimmaturityinsensitivityto

coher-entmotionattheslowestspeedtested(0.1◦/s)compared

to faster speeds of 0.9 and 5◦/s Also, in a speed

dis-criminationtask,Manningetal.(2012)reporteda more

gradualdevelopmentofthresholdsforslow(1.5◦/s)than

fast(6◦/s)speeds.However,Hadadetal.(2011)didnot

finddifferentratesofdevelopmentformotioncoherence

thresholdsmeasuredwithrandomdotstimulimovingat

4◦/sand18◦/s.Together,thisresearchsuggeststhatmotion

processing forintermediate and fastspeedsmay follow

similarratesofdevelopment,butthatprocessingofmuch

slower speeds (e.g., 0.1 and 1.5◦/s) may develop more

slowly

Globalmotionprocessingabilitiesinchildhoodare

gen-erallythoughttobelimited bypoorintegrationoflocal

motioncuesoverspace(e.g.,Ahmedetal.,2005;Hadad

et al., 2011; Manning et al., 2012) Such integration is

believed to occur in higher-order areas of the motion

processing hierarchy, suchas in area MT/V5 (Born and

Tootell,1992;Brittenetal.,1992).Yetperformanceontasks

traditionallyusedtoassessglobalmotionprocessing(i.e.,

motioncoherenceparadigms;NewsomeandParé,1988)

isnotlimitedsolelybyglobalintegration.Suchtasksare

likelylimitednotonlybyanobserver’sabilitytoglobally

poolthemotionofindividualdotsacrossspace,butalsoby

theirabilitytoestimatethelocalmotiondirectionofeach

dot(BarlowandTripathy,1997),andbytheirabilityto

seg-mentthesignaldotsfromthemaskingnoise(Dakinetal.,

2005;Tibberetal.,2014;Websteretal.,2011)

Increasedneuralvariabilitywouldleadtoimprecisionin

estimatingindividualdotdirections,which,whenpooled,

couldleadtoelevatedmotioncoherencethresholds.This

neural variability has been termed‘internal noise’, and

hasmanypotentialsources,includingphotonnoise,

vari-ability in the firingof action potentials, and variability

in synaptic transmission (Faisal et al., 2008) Through development,neuronsinareaV1undergoextensive synap-tic pruning(Gareyand de Courten, 1983; Huttenlocher

etal.,1982;HuttenlocheranddeCourten,1987),andthe bandwidthsofdirection-selectivecellsreducewithage(at leastintheprimatebrain,Hattaetal.,1998).Itis possi-blethatsuchdevelopmentalchangesmightbemanifestas reducedinternalnoisewithage

The traditional motion coherence paradigm cannot distinguishbetweenlocalandgloballimitstomotion per-ception and has hence obscured our understanding of whatlimitsglobalmotionprocessingduringdevelopment (andinavarietyofneurodevelopmentaldisorders;Dakin andFrith,2005).Toaddressthisissue,thecurrentstudy usedtheequivalentnoiseparadigm(Barlow,1956;Pelli,

1990)to determine whether local or global processing limitsmotioncoherencesensitivityindevelopment.The equivalentnoiseparadigmisbasedoncomparinghuman performancetothat ofanidealobserverthat islimited bothbyadditiveinternalnoiseandbyhowcompletelyit samplestheinformationavailablefromthestimulus(Pelli,

1990).Whenequivalentnoiseanalysisisappliedto direc-tiondiscrimination(Dakinetal.,2005),internalnoisemaps onto theprecisionwithwhich individual motion direc-tionsareestimatedand samplingrepresentsanestimate

of the effectivenumber of local motiondirections that aregloballypooled(oraveraged).Whereasmotion coher-encestimulicontainbothsignaldotsandrandomlymoving noise dots,equivalentnoise stimuli containdotswhose directions (onany one trial)are sampledfrom a single Gaussian distribution(Dakinet al., 2005).The standard deviationofthis distributionisvariedacrossconditions,

inordertomanipulatethelevelofstimulusvariability(or

‘externalnoise’;seeFig.1A)

Intheequivalentnoisetask,theobserverisaskedto dis-criminatethemeandirectionofdots,andtheperformance measureisthesmallestdifferenceindirectionfromafixed referencedirection(e.g.,upwards)thatobserverscan reli-ablyreport.Withnodirectionalvariance (i.e.,whenthe standarddeviationis0◦andallelementsmoveinthesame direction),theobserver’sperformanceislimitedbothby internalnoiseandsampling.Consequently,smallamounts

ofextraexternalnoisehavelittleeffectonthresholds,as

itisswampedbytheobserver’sowninternalnoise How-ever,asthelevel ofexternal noise isincreased,a point

is reachedwhere theexternal noise exceeds the inter-nalnoiseinherentinthesystem,andthresholdsstartto increase withtheadditionoffurtherexternal noise An equivalentnoisefunctioncanbefittothesedatatoderive estimatesoftheindividual’sinternalnoiseandsampling (seeFig.1A)

Asthresholdsaremeasuredacrossarangeofexternal noiselevels,theequivalentnoisemethodtypicallyrequires severalthousandtrials,makingit unsuitablefor investi-gatingthevisualabilitiesofchildren,whomaygetbored andbecomeinattentive.However,amoreefficient equiv-alentnoiseprocedurehasbeendeveloped,whichprovides reliableestimatesofinternalnoiseandsamplinginfewer than100trials(Tibberetal.,2014).Inthisnovelmethod, twohighlyinformativepointsontheequivalentnoise func-tionareprobed(seegreyline,Fig.1B).Inonecondition

Trang 3

Fig 1. (A) Equivalent noise functions relating direction discrimination thresholds to the standard deviation of dot directions (i.e., external noise) Lower sampling is represented by an equivalent noise function that is shifted vertically upwards, whilst higher levels of internal noise require more external noise

to be added before thresholds increase (B) The black circles and curve represent the standard equivalent noise paradigm where direction discrimination thresholds are measured at multiple levels of external noise Large standard deviations of dot directions reflect high external noise in the stimulus The grey circles and curve are derived using a rapid version of the equivalent noise paradigm, which measures performance at two maximally informative noise levels In the ‘no noise’ condition, there is no external noise (i.e., the standard deviation of dot directions is 0 ◦ ) and the threshold is taken as the finest direction discrimination possible In the ‘high noise’ condition, we measure the maximum noise that can be tolerated when the observer is judging if the pattern is moving either +45◦or −45 ◦ of vertical (C) Example of a stimulus in the ‘low noise’ condition, where the mean direction of dots is +4◦, and the standard deviation of directions is 0◦ (D) Example of a stimulus in the ‘high noise’ condition, where the mean direction of dots is +45◦, and the standard deviation of dot directions is 45 ◦ Arrows are provided for illustrative purposes only, to represent the direction of motion.

(‘nonoise’, Fig.1C),thestandard deviationofdot

direc-tionsis0◦,andanadaptivestaircaseprocedureisusedto

estimatethefinestdirectiondiscriminationpossible.Inthe

othercondition(‘highnoise’,Fig.1D),anadaptivestaircase

procedureestimateshowmuchdirectionalvariabilitycan

betoleratedwhilediscriminatingalarge(±45◦)directional

offset.Asthesethresholdshaveconfidenceintervalsthatlie

inorthogonalplanes,thefitoftheequivalentnoise

func-tionisefficientlyconstrainedtoprovidereliableestimates

ofinternalnoiseandsampling

Here,weusedTibberetal.’srapidversionofthe

equiv-alent noise direction integration paradigm alongside a

traditionalmotioncoherencetasktoinvestigatethe

fac-torslimitingthedevelopmentofglobalmotionprocessing

Thesemethodsallowedustoinvestigate(1)how

inter-nal noise and sampling develop, and (2) the extent to

whichchangesinthesefactorsimpactuponacommonly

usedmeasureofglobalmotionprocessing,namelymotion

coherencethresholds Due to the possibility of distinct

developmentaltrajectoriesfordifferentspeeds(Hayward

etal.,2011;Manningetal.,2012),equivalentnoiseand

motioncoherencetaskswerepresentedattwostimulus

speeds:slow(1.5◦/s)andfast(6◦/s)

Itiscommonlyassumedthatmotioncoherence

thresh-olds are limited by poor integration of local motion

information(e.g.,Hadadetal.,2011).Wetherefore

hypo-thesised that sampling would increase with age and

that this would contribute to age-related reductions in motioncoherencethresholds.Derivinghypothesesabout thedevelopmentofinternalnoisewaslessstraightforward Indeed,someresearchershavenotedthatchildren have hightrial-to-trialbehaviouralvariabilitywhichdecreases with age (e.g., Williams et al., 2005), where higher behavioural variability isthoughtto reflect higher neu-ronalvariability(i.e.,noise;Foxetal.,2007,butseealso

Beck et al., 2012) Additionally, Skoczenski and Norcia (1998)measuredinternalnoiseininfantsusingan equiv-alentnoisetechniquewithvisuallyevokedpotential(VEP) responsesandreportedthathighlevelsofinternalnoise

ininfancylimitedcontrastsensitivity.Similarly,Bussetal (2006)demonstratedincreasedlevelsofinternalnoisein childrenaged5–10yearscomparedtoadults,withchildren beinglesssusceptibletotheeffectsofexternalnoise (rov-ingintensities)inanauditoryintensitydiscriminationtask Alternatively,somehavesuggestedthatneuronal variabil-ityin factincreaseswithagefrom8years toadulthood,

as measured by trial-by-trial EEG variability (McIntosh

etal.,2008).Evidently,internalnoisehasbeenmeasured

in arange ofdifferentways andit isnot yetclearhow thesemeasuresrelatetoeachother.Inthecurrentstudy,

wethereforeaimedtoinvestigatehowinternalnoiseand samplingchangethroughchildhoodforadirection integra-tiontaskandtodeterminewhethersuchchangeslimitthe developmentofmotioncoherenceperception

Trang 4

2 Materials and methods

2.1 Participants

Fivegroupsofparticipantsweretested,with21

5-year-olds(M=5years;4months,range4;10–5;10,14females),

277-year-olds(M=7years;3months,range6;7–7;10,

11females),259-year-olds(M=9years;2months,range

8;8–9;9, 11females),20 11-year-olds (M=11 years;3

months, range 10; 8–11; 9, 14 females) and 30 adults

(M=26years;9months,range21;5–35;10,15females)

includedinthefinaldataset.Childrenwererecruitedfrom

schoolsintheSouthEastofEngland.Normalor

corrected-to-normalvisualacuitywasconfirmedbybinoculartesting

withletteracuitytestsusingopticalcorrectionswhere

nec-essary.Normalacuity wasdefined asa binocularacuity

of6/9orbetterfor5-and7-year-olds(becauseacuityis

stillmaturinginthisagerange;AdamsandCourage,2002;

Ellembergetal.,1999)and6/6orbetterfor9-and

11-year-oldsandadults

Anadditionalnine5-year-oldswereexcludedfromthe

dataset,with onechild failing to passthe visualacuity

screening,onefailingtoreachcriterion(seeSection2.3.1),

three not performing significantly above chance in the

catchtrials(seeSection2.6.2)andfourobtainingmotion

coherencethresholdsabove100%,indicatinganinability

toperformthetask.Oneadditional7-year-oldcouldnot

completethemotioncoherencetask.Anadditionaltwo

9-year-olds,one11-year-oldandoneadultwereexcluded

fromthedatasetduetodiagnosesofdevelopmental

con-ditions

2.2 Apparatusandstimuli

ThestimuliwerepresentedusingMATLAB(The

Math-worksLtd.)usingelementsofthePsychophysicsToolbox

software(Brainard,1997;Kleineretal.,2007;Pelli,1997)

Stimuli were displayed on a Dell Precision M4600

lap-top at a frame rate of 60Hz and a pixel resolution of

1366×768

A yellow-bordered circular aperture (diameter=15◦) andanchor-shapedfixationpoint(0.57×0.57◦)were pre-sented against a grey background witha luminance of

30cd/m2(seeFig.2).Twosmalleryellow-bordered circu-larapertures(diameter=6.12◦)werepresentedtotheleft andrightofthis,servingasreferencepointsforthe repor-tingofmotiondirection.Intheequivalentnoisetask,the leftandrightapertureswerepresentedinthetopcorners

ofthescreenandcontainedimagesofredandgreenreefs, respectively(seeFig.2A).Inthemotioncoherencetask,the leftandrightapertureswerepresentedhalfwaydownthe screen,containingimagesofredandgreenrocks, respec-tively(seeFig.2B)

Thestimuliwerecomprisedof100randomlypositioned white dots(58.7cd/m2), each with a diameterof 0.44◦, driftingfor400mswithinthecentralaperture.Dot pos-itionswereupdatedevery3frameswithdisplacementsof 0.075◦ and0.3◦intheslow(1.5◦/s)andfast(6◦/s) condi-tions,respectively.Dotswereallowedtooverlapandwere notlimitedintheirlifetime

2.3 Procedure Participantscompletedanequivalentnoisetaskanda motioncoherencetaskineach oftwospeedconditions: slow(1.5◦/s)andfast(6◦/s).Intheequivalentnoisetask, dotdirections wererandomlysampledfroma wrapped normaldistributionwitha specifiedmeanandstandard deviation.Theequivalentnoisetaskconsistedoftwo inter-leavedconditionsthatprobedtwoinformativepointson theequivalentnoisefunction toconstrainthefitof the model(seeFig.1B).Inthe‘nonoise’condition,thestandard deviationof dotdirections wasfixedat0◦ (i.e.,alldots movedinthesamedirection),whilethemeandirection

ofthedotswasvaried(leftwardorrightwardofvertical)

tofindthefinestdirectionthatcouldbediscriminated84%

ofthetimeintheabsenceofstimulusnoise (correspond-ingtothemeanplusonestandarddeviationinacumulative normaldistribution).Inthe‘highnoise’condition,themean directionofdotswasfixedat45◦leftwardsorrightwards

Fig 2.Schematic representation of stimuli presented in the ‘high noise’ condition of the equivalent noise task (A) and the motion coherence task (B) The

Trang 5

dotdirectionswasvariedtofindthemaximumlevelof

noisethatcouldbetoleratedwhilstidentifyingthesignal

directionwith84%accuracy

Theequivalentnoisetaskwaspresentedas“TheHungry

FishGame”.Participantsjudgedwhetherashoalof“fish”

was“swimming”towardsthered(left)orgreen(right)reef

tofindtheirfood.Childrenweretoldthatsometimesthe

fishallmovedinthesamedirection(‘nonoise’)and

some-timesthefishmovedindifferentdirections(‘highnoise’),in

whichcasetheyhadtodeterminetheoverall(i.e.,average)

direction.Toaidmotivation,childrenweretoldthatthey

werecompetingagainstacartooncharacter,“ScubaSam”

In the motion coherence task, a proportion of dots

movedcoherently in asingle direction(90◦ leftwardor

rightwardofvertical)whiletheremainingdotsmovedin

randomdirections.Thetaskwaspresentedwithinthe

con-textof“TheSharkAttackGame”.Participantswereaskedto

judgewhethertheshoalof“fish”was“swimming”towards

thered(left)orgreen(right)rockstohidefromthe“shark”

Childrenweretoldthatthe“fish” sometimes“panicked”

whentheysawthe“shark”,causingthemtogoindifferent

directions.Toenhancemotivation,childrenweretoldthat

theywerecompetingagainstthe“shark”

Each equivalent noise and motion coherence task

consistedofthreelevels:acombineddemonstrationand

criterionphase(‘level1’),apracticephase(‘level2’),anda

thresholdestimationphase(‘level3’).Inalllevelsinboth

tasks,direction(leftwardorrightwardofvertical)was

ran-domisedoneachtrial

2.3.1 Demonstrationandcriterionphase

Theexperimenterexplainedeachtasktoparticipants

withinthe context of four demonstrationtrials, two of

which were designed to be ‘easy’, and two of which

were‘slightly harder’.In theequivalentnoise task, two

of thetrials demonstrated the ‘no noise’ condition,and

two demonstrated the‘high noise’ condition.Next,

par-ticipantswere presented withup to 20 criterion trials

Intheequivalentnoisetask,‘nonoise’ stimuliwere

pre-sentedwitha directionof 45◦ leftwardorrightward of

vertical.Inthemotioncoherencetask,dotsmovedwith

100%coherence90◦leftwardorrightwardofvertical

Par-ticipantswhofailedtoreachacriterionoffourconsecutive

correctresponseswithin20trialsweregivenashort

ver-sionofthetaskandexcludedfromanalysis(n=1).Children

respondedeitherverballyorbypointing,withthe

experi-menterrelayingtheresponsetothecomputer.Visualand

verbalfeedbackandencouragementwereprovided

2.3.2 Practicephase

Eightpracticetrialswerepresentedinafixedorderfor

eachtaskwithincreasingdifficulty.Intheequivalentnoise

task,four‘nonoise’stimuliandfour‘highnoise’stimuliwere

presentedinalternatingorder.Participantsreceived

feed-backasbefore,buttherewasnocriterionforproceedingto

thenextphase

2.3.3 Thresholdestimationphase

Boththeequivalentnoiseandmotioncoherencetasks

employedtheQUESTadaptivestaircasemethod(Watson

and Pelli,1983).Intheequivalentnoisetask, two stair-cases(75trialseach)wereinterleavedforeachofthe‘no noise’and‘highnoise’conditions.Inthe‘nonoise’ condi-tion,theQUESTfunctiontrackedthebasicdirectionoffset thresholdintheabsenceofnoise.Inthe‘highnoise’ condi-tion,themeandirectionofmotionwassetto±45◦ and

QUEST tracked the maximum level of noise that could

be tolerated whilst discriminating the mean direction

Anadditional15catchtrialswereinterleavedrandomly, presenting stimuli identical to those used in the crite-rionphase.Thisyielded165trialsintotalforeachspeed condition

Inthemotioncoherencetask,asingleQUESTstaircase

of75trialstrackedtheminimumcoherencelevelrequired foraccurate(84%correct)directiondiscrimination.Asin the equivalent noise task, there were an additional 15 catch trials, which presented stimuli used in the crite-rion phase This resulted in 90 trials in total for each speed condition.Trials weredivided intofourblocks of equal lengthfor each conditionof eachtask Whenthe end of a blockwasreached, participantswereshown a simulated graph of the “points” they and their “oppo-nent”(“ScubaSam” or the“shark”)had attained These pointswererandomlyjitteredaroundafixedsetofvalues

tominimiserewardandmotivationeffects onthreshold estimates

2.4 Eyetracking

Toestablishwhetherdevelopmentaldifferencesintask performancecouldbeaccountedforbydifferencesin abil-itytomaintainfixation,weusedaTobiiX2-30Compact eyetrackermountedontothescreentocollectfixationdata forasubsetofparticipants,including12five-year-olds,17 seven-year-olds,11nine-year-olds,911-year-oldsand10 adults.Afive-pointcalibrationprocedurewasconducted beforetheintroductoryphaseandfixationdatawere sam-pledatarateof40Hzduringstimuluspresentationinthe thresholdestimationphase

2.5 Generalprocedure TheprocedurewasapprovedbytheInstituteof Edu-cation’s Faculty Research Ethics Committee All adult participantsand parentsofchild participantsgavetheir informed consent.Childrenprovided verbalassent Par-ticipantswereseenindividuallyfortwo sessionslasting approximately 25min, each consisting of one equiva-lent noiseand onemotioncoherencetask.Theorder of presentation of conditions was counterbalanced across participants Participants were seated in a dimly lit

binocularly using a chin-rest They were instructed to maintain central fixation throughout stimulus presen-tation, which the experimenter monitored, providing reminders to maintain fixation and only initiating tri-als when the participant was attending Participants were each given a ‘Submarine Log Book’ with which they recorded their progress through the experimental session

Trang 6

2.6 Dataanalysis

2.6.1 Equivalentnoiseanalysis

Theequivalentnoisemodeldescribeschangesin

direc-tion discrimination threshold as a function of external

noise:

2

2

ext

where2

obsistheobserver’sthreshold,2

intisadditive inter-nalnoise,2

extistheexternalnoiseaddedtothestimulus,

andnsampistheeffectivenumberofsamplesusedto

cal-culatethemeandirectionofthestimulus.Thisapproach

exploitsadditivityofvariance,wherebyinternalnoiseand

externalnoisecontributeindependentlytoanobserver’s

directiondiscriminationthreshold

Theequivalentnoisetaskyieldedtwo thresholds:(a)

thefinestdirectiondiscriminationpossiblewithno

stim-ulus noise (‘no noise’ condition), and (b) the maximum

levelofnoisethatcouldbetoleratedwhilst

discriminat-ingalargesignaloffsetof45◦ (‘highnoise’condition).By

running Monte Carlosimulations of a model observer’s

performance acrossa rangeofinternalnoise and

samp-linglevels, Bexet al.have shown that – assumingthat

a participant’s internal noise is negligibleat high noise

levels–sampling(nsamp)canbeestimatedfromalinear

transformationoftheirmaximumtolerablenoisethreshold

(MTN):

nsamp=exp(0.000121∗MTN2+0.0357∗MTN−1.8093)

(2)

Asperformanceatlowlevelsofexternalnoiseis

deter-minedbothbyinternalnoiseandsampling,itispossibleto

usetheestimateofnsamptocomputethelevelofinternal

noise,byrearrangingEq.(1).Thus,whenexternalnoiseis

zero(2

2

Thisapproachassumesthatobservers donotchange

theirsampling(ormoregenerally,theirstrategy)asa

func-tionofexternalnoiselevel.Consistentwiththisview,the

equivalentnoisefunctionhasbeenshowntofitdirection

discriminationdataover awide range ofexternal noise

levels(directionalvariability),undervaryingstimulus

con-ditions(Dakinetal.,2005).Notethatthisapproachdoes

notassumethatobservers arenecessarilyaveragingdot

directionsinthewaythemodeldoestomakeperceptual

judgements.Nomatterhowobserversperformthetaskthe

modelwillreturntheeffectivenumberofsamplesthatare

averaged–thatistosaythattheobserverisactingasifthey

wereaveragingacertainnumberofdots.Thusallnoiseand

samplingestimatesquotedarenecessarilyeffectivevalues

sincewecannotknowtheobserver’sunderlyingstrategy

forperformingthetask

2.6.2 Datascreeningandtransformation

Alapseratewascalculatedastheproportionof

incor-rect responses to catch trials for each participant for

each condition for each task A binomial test revealed

thatparticipantsrespondingincorrectlyon4ormoreof thecatchtrials werenotperformingsignificantly above chance.Threefive-year-oldswerethereforeexcludedfrom analyses(seeSection2.1)

Analysisofvariance(ANOVA)showedthatlapserates differed significantly across age groups, F(4,118)=9.26,

p<.01,2

p=.24(5-year-olds:M=.04,SD=.06;7-year-olds:

M=.02; SD=.04, 9-year-olds: M=.01, SD=.03; 11-year-olds:M=.01, SD=.03;adults, M<.01, SD=.01).Post hoc Dunnettt-testscomparingeachoftheagegroupswiththe adultgroupsrevealedthat5-year-oldsand7-year-oldshad significantlyhigherlapseratesthanadults (5-year-olds:

p<.01;7-year-olds:p<.01),whereasthe9-and 11-year-oldsdidnot differfrom theadultgroup(p>.05).There wasnomaineffectoftask(p=.45),althoughhigherlapse rateswerefoundfortheslowspeedconditions(M=.02,

SD=.05)thanthefastspeedconditions(M=.01,SD=.03), F(1,118)=15.40,p<.01,2

p=.12.Nointeractionswere sig-nificant(ps>.05)

Toensurethatanyage-relatedand/orspeed-related dif-ferencesininternalnoise,samplingormotioncoherence thresholdswerenotaby-productofdifferencesin atten-tion,anidealobservermodelwasrunassumingdifferent levelsoflapserate.MonteCarlosimulationsallowedusto modeltheeffectofdifferinglapseratesonthresholds.We averagedacrosstaskstogetalapserateforeachobserverin eachspeedcondition,andthencorrectedthethresholdsfor eachobserveraccordingtotheirlapserateforeachspeed condition,basedonthesimulationresults

Next,theinternalnoise,samplingandmotion coher-encethreshold estimates in each speed condition were assessedforskewnessandkurtosis.Allmeasuresshowed significantpositiveskew(ps<.05)andthemajorityshowed significantkurtosis(ps<.05).Consequently,alldatawere log-transformed.Thedatawerethenscreenedforoutliers lyingmorethanthreezscoresfromthemeanforeachage groupineachspeedcondition.Nooutlierswerefoundin motioncoherencethresholds,internalnoiseorsampling estimates.Alloftheanalysesreportedbelowwere con-ductedwithlog-transformed,lapse-correctedvalues 2.6.3 Fixationanalysis

Rawfixationdatawere(x,y)coordinatessampledduring stimuluspresentationineachtrialofthethreshold estima-tionphaseforleftandrighteyepositionsrelativetothe screen’sactivedisplayarea.Thedatawereinitiallyfiltered accordingtoavaliditycodefrom0(signifyingtheeyewas definitelyfound)to4(signifyingtheeyewasnotfound) Allsampleswithvaliditycodes of2orhigherwere dis-carded(TobiiTechnology,2013).The(x,y)coordinateswere thenaveragedacrosstheleftandrighteyeforanalysis.A measureoffixationstabilitywasderivedbypoolingthe standarddeviationsoffixationlocationsinxandy dimen-sions.Thestandarddeviationswerethenlog-transformed

tominimisetheeffectsofskewnessandkurtosis

3.1 Age-relatedchangesininternalnoise Levelsofinternalnoisereducedwithage,with 5-year-oldshavingmeanlevelsof9.62◦and9.69◦intheslowand

Trang 7

Fig 3. Individual values for internal noise (A), sampling (B) and motion coherence thresholds (C) for slow (1.5◦/s) (open red circles) and fast (6◦/s) (filled blue circles) conditions as a function of age Red dashed and blue solid lines represent the line of best fit for the slow and fast conditions, respectively (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)

Trang 8

fastconditions,respectively,whichreducedto6.72◦ and

4.80◦intheadultgroup.Tocharacterisetherateof

devel-opmentalchangesinestimatedinternalnoise,loginternal

noise values wereplotted asa function of logage and

fit witha straight line(Fig.3A).Wethen comparedthe

developmentaltrajectoriesforslowandfastspeedsusing

theANCOVAmethodoutlinedbyThomasetal.(2009).In

thismethod,within-subjectseffectsareinitiallyexamined

usinganANOVAbeforeassessingage-relatedchangesby

addingacovariate(aswithin-subjectseffectsaremasked

whenabetween-subjectscovariateisadded;Delaneyand

Maxwell,1981; Thomas et al., 2009) AninitialANOVA

withspeedcondition(slow,fast)asawithin-subjectsfactor

revealedthatsignificantlyhigherlevelsofloginternalnoise

werefoundintheslow(M=.87,SD=.24)thanthefast

con-dition(M=.79,SD=.25),F(1,122)=12.24,p<.01,2

Next,anANCOVAwasconductedbyaddinglogageintothe

modelasacovariate.Overall,loginternalnoisereduced

sig-nificantlywithage,F(1,121)=13.42,p<.01,2

p=.10.Also, there wasasignificantinteractionbetweenlogageand

speedcondition,F(1,121)=4.76,p=.03,2

p=.04, indicat-ingasignificantlysteeperrateofdevelopmentinthefast

conditionthantheslowcondition

Dunnett t-tests (corrected for multiplecomparisons)

wereconductedtodeterminewhenadult-likelevelsoflog

internalnoisewerereachedforslowandfastspeed

condi-tions.Intheslowcondition,5-year-oldshadsignificantly

higherloginternalnoisethanadults(p=.02)whereas7-,

9- and11-year-oldshad adult-likelevelsofloginternal

noise(ps>.05)(5-year-olds:M=.98,SD=.25;7-year-olds:

M=.86, SD=.27; 9-year-olds: M=.87, SD=.23;

11-year-olds:M=.83,SD=.22;adults:M=.83,SD=.23).Similarly,

inthefastcondition,5-year-oldshadhigherloginternal

noisethanadults(p<.01)whereastheolderagegroupsdid

not (ps>.05)(5-year-olds: M=.99, SD=.28; 7-year-olds:

M=.78, SD=.23; 9-year-olds: M=.80, SD=.20;

11-year-olds:M=.78;SD=.21;adults:M=.68,SD=.26)

3.2 Age-relatedchangesinsampling

Next,weinvestigatedage-relatedchangesinsampling

AsshowninFig.3B,samplingincreasedfrom0.51atage

5to1.47inadultsin theslowcondition,andfrom0.98

to1.85inthefastcondition.WerepeatedtheANOVAand

ANCOVAanalysesusinglog-transformedlevelsofsampling

asthedependentvariable.Higherlevelsoflogsampling

wereobtainedinthefastcondition(M=−.06;SD=.34)than

intheslowcondition(M=.13;SD=.34),F(1,122)=39.12,

p<.01,2

p=.24.Whenlogagewasaddedintothemodel

asa covariate,it wasfoundthatlogsamplingincreased

acrossdevelopment,F(1,121)=23.32,p<.01,2

predicted However, there was no interaction between

speed conditionandlogage,F(1,121)=1.88,p=.17,

sug-gesting a similar rate of development in slow and fast

conditions

Dunnettt-testsrevealedthatallchildgroupshadlower

logsamplingcomparedtoadults(M=.17,SD=.35)inthe

slowcondition(5-year-olds:M=−.29,SD=.28;p<.01;

7-year-olds:M=−.06,SD=.35;p<.01;9-year-olds:M=−.15,

SD=.28;p<.01;11-year-olds:M=−.06,SD=.28;p=.01)

Inthefastcondition,5-year-olds(M=−.01,SD=.28)and

7-year-olds (M=−.01; SD=.25) had lower log sampling thanadults(M=.27;SD=.40)(ps<.01)whereas9-year-olds (M=.14,SD=.33)and 11-year-olds(M=.23,SD=.30)did notdiffersignificantlyfromadults(ps>.05)

3.3 Age-relatedchangesinmotioncoherencethresholds Whereas5-year-oldsrequired,onaverage,47% coher-ent motion in both the slow and fast conditions to reliably report the directionof motion, adults required only 34% and 26% coherent motion in the slow and fast conditions, respectively The ANOVA and ANCOVA analyses were repeated to characterise developmental changes in log motion coherence thresholds (Fig 3C) Higher logmotioncoherence thresholdswere foundin theslowcondition(M=−.41,SD=.16)thanthefast con-dition(M=−.51,SD=.21),F(1,122)=37.18,p<.01,2

Thresholdsdecreasedwithlogage,F(1,121)=20.50,p<.01,

2

p=.14,buttherewasnosignificantinteractionbetween speedconditionandlogage,F(1,121)=2.73,p=.10, indi-catingthatsensitivitydevelopedatasimilarrateforslow andfastspeeds

Five-year-oldsand7-year-oldshadsignificantlyhigher log thresholds than adults in both the slow and fast conditions(ps<.01),whereas9-and11-year-oldsshowed adult-like levels of performance (ps>.05) (5-year-olds:

Mslow=−.33,SDslow=.18,Mfast=−.32,SDfast=.17; 7-year-olds: Mslow=−.35, SDslow=.14, Mfast=−.44, SDfast=.17; 9-year-olds: Mslow=−.42, SDslow=.15, Mfast=−.58,

SDfast=.22; 11-year-olds: Mslow=−.48, SDslow=.14,

Mfast=−.63,SDfast=.16;adults:Mslow=−.47,SDslow=.16,

Mfast=−.58,SDfast=.17)

3.4 Relationshipbetweenequivalentnoisemeasuresand motioncoherencethresholds

Our results show that internal noise reduces, and samplingincreases,throughdevelopment,whilemotion coherencethresholdsdecrease.Nextwesoughtto inves-tigatewhetherincreasingsensitivitytocoherentmotion

is driven either by internal noise or sampling, or a combinationofboth.Correlationanalysesincludingall par-ticipantsrevealednorelationshipbetweeninternalnoise and motioncoherencethresholdsin eitherslow, r=.03,

df=122,p=.77,orfast,r=.08,df=122,p=.36,conditions However,samplingwasnegativelycorrelatedwithmotion coherencethresholdsinbothslow,r=−.35,df=122,p<.01, andfast,r=−.34,df=122,p<.01,conditions

Webuiltahierarchicalregressionmodelonlogmotion coherencethresholdsforeachspeedcondition,withlogage addedintothemodelfirst,followedbysamplingand inter-nalnoiseaddedinastepwisemanner(seeTable1).Inboth slowand fastconditions,logagesignificantly predicted motioncoherencethresholdsinthefirststepofthemodel Whensamplingandinternalnoisewereaddedintothe sec-ondstepofthemodel,ageremainedasignificantpredictor

ofmotioncoherencethresholds,andsamplingwasalsoa significantpredictorinbothslowandfastconditions Inter-nalnoise,however,failedtosignificantlypredictcoherence thresholdsforeitherspeedcondition(slow,ˇ=.14,p=.16,

orfast,ˇ=.08,p=.41),andwasthereforeexcludedfrom

Trang 9

Table 1

Hierarchical regression analyses on log motion coherence thresholds in

the slow (1.5 ◦ /s) and fast (6 ◦ /s) conditions.

Step 1

Step 2

Log age −0.12 0.06 −.18 * −0.24 0.07 −.29 **

Log sampling −0.13 0.04 −.28 ** −0.16 0.05 −.26 **

Note:

In the slow condition, R 2 = 09, p < 01 for Step 1; R 2 = 06, p < 01 for Step

2 In the fast condition, R 2 = 13, p < 01 for Step 1; R 2 = 06, p < 01 for Step

2.

* p < 05.

** p < 01.

themodelinbothspeedconditions.Step2ofthemodel,

withbothlogageandsampling,wasasignificantlybetter

model than Step 1 of the model in both speed

condi-tions(seeTable1).Theresultingmodelwithlogageand

logsamplingsignificantlypredictedlogmotioncoherence

thresholdsinbothslow(F(2,120)=10.63,p<.01)andfast

(F(2,120)=14.32,p<.01)conditions

3.5 Fixationanalysis

Next,weinvestigatedwhethertherewereage-related

changesin theabilitytomaintainfixationand whether

thesewererelatedtotaskperformance.Thestandard

devi-ationofparticipants’eyepositionsforeachtaskisshown

inFig.4.ApreliminaryANOVAonstandarddeviationsin

theequivalentnoisetaskrevealednomaineffectofnoise

condition(‘nonoise’,‘highnoise’)andnointeractionswith

agegrouporspeedcondition,andsothisfactorwasnot analysedfurther

AmixedANOVAwasconductedonthestandard devi-ationswithspeed(1.5◦/s,6◦/s)andtask(equivalentnoise, motioncoherence)aswithin-participantsfactorsandage group(5-,7-,9-and11-year-oldsandadults)asa between-participantsfactor.Therewasnomaineffectofstimulus speed, F(1,54)=1.34, p=.25 However, higher standard deviations(i.e.,reducedstability)werefoundinthe equiv-alent noise task (M=−1.08, SD=.28) than the motion coherencetask(M=−1.22,SD=.28),F(1,54)=52.47,p<.01,

2

p=.49 There was a significant main effect of age, F(4,54)=4.08, p<.01, 2

p=.23 Dunnett t-tests revealed that 5-year-oldshadsignificantly largerstandard devia-tions(M=−.92,SD=.27)thanadults(M=−1.27,SD=.28),

p<.01,whereastheolderagegroupswerenotsignificantly differenttoadults(7-year-olds:M=−1.14,SD=.27; 9-year-olds:M=−1.26,SD=.26;11-year-olds:M=−1.19,SD=.23;

ps>.05).Nosignificantinteractionswerefoundbetween task,speedconditionandgroup(ps>.05)

Havingfoundthattheyoungestchildrenhavelessstable fixations than older participants, we sought to investi-gate whetherthesedifferencesrelated tointernal noise and sampling.Giventhathigher levelsofinternal noise andlowersamplingarefoundintheslowcondition (Sec-tions3.1and3.2),weconductedseparateanalysesforeach speedcondition.Intheslowcondition,fixationstandard deviationwasrelatedtointernalnoiseestimates,r=.28,

df=58,p=.04,withlowerfixationstandarddeviations(i.e., morestablefixations)beingassociatedwithlower inter-nalnoise Therewas,however,norelationshipbetween fixationstandarddeviationandsampling,r=−.14,df=58,

p=.29.Similarly, inthefastcondition,fixation standard deviation was related to internal noise, r=.30, df=58,

p=.02,butnotsampling,r=−.05,df=58,p=.72.Finally,we investigatedtherelationshipbetweenfixationstabilityand

Fig 4.Standard deviations of eye positions in equivalent noise tasks (left panel) and motion coherence tasks (right panel) for slow (1.5◦/s) and fast (6◦/s) speed conditions Circles show individual performance (slow: open circles; fast: filled circles) and lines represent mean performance for each age group (slow: red dotted line; fast: blue solid line) Standard deviations were log-transformed for analysis (For interpretation of the references to color in this

Trang 10

thresh-oldswerenotrelatedtostandarddeviationofeyepositions

ineitherslow,r=.14,df=58,p=.31,orfast,r=.17,df=58,

p=.20,conditions

Thisstudypresentedanequivalentnoisemotion

inte-grationtaskalongsideatraditionalmotioncoherencetask

tochildrenaged5,7,9and11years andadultsfortwo

speed conditions (slow: 1.5◦/s; fast: 6◦/s) Thismethod

allowed us to characteriseboth age-related changes in

internal noise and sampling and the mechanisms

sup-porting coherent motion processing While there was

considerableindividual variability, wefound that

inter-nalnoiseestimatesreducethroughchildhood,reflecting

improvedlocalprocessing,andthat thisisaccompanied

byanincreaseinthenumberofsamplesthechildcanuse

toestimateglobalmotion.Notethattheeffectivenumber

ofsamplescanalsobethoughtofasameasureof

multi-plicativenoisebeingaddedtoallestimatesinthepooling

process(i.e.,‘globalnoise’;Dakinetal.,2005).Although

levelsofinternalnoisereducedwithage,thesedidnot

pre-dictmotioncoherencethresholds.Instead,developmental

increases in motioncoherence sensitivity appear to be

drivensolelybyage-relatedincreasesinsampling

Overall,higherlevelsofinternalnoiseandlower

samp-lingwere foundin theslow(1.5◦/s)condition thanthe

fast(6◦/s)condition,whichmightreflectdistinct

speed-tunedmotionprocessingsystems (e.g.,Thompsonetal.,

2006).Generallypoorer performancemight bea

conse-quenceoffewerneuronstunedtoslowspeedsthanfast

speeds,asfoundintheprimatebrain(Hadadetal.,2011;

LiuandNewsome,2003).Wewereparticularlyinterested,

however, in how internal noise and sampling changed

withage,and how theseage-relatedeffects might vary

betweenspeed conditions.Internalnoise levelsreduced

more gradually in the slow (1.5◦/s) condition than the

fast(6◦/s)condition,whereassamplingfollowedasimilar

rate of developmentfor slow and faststimuli

Further-more,samplingappearedtofollowamoreprotractedrate

ofdevelopmentthanthatofinternalnoise.Internalnoise

reachedadult-likelevelsbyapproximately7yearsofage,

while sampling reachedadult-likelevels at a later age

Indeed,samplingwasadult-likeby9yearsinthefast

con-dition,butwasnotyetadult-likeby11yearsintheslow

condition

OurresultscomplementarecentstudybyBogfjellmo

et al (2014),which found increasedsamplingof

direc-tioninformationbetweentheagesof6and17yearsfor

stimulusspeedsof2.8◦/sand9.8◦/s,whilelevelsof

inter-nal noise remained stable Taken together, the current

resultsandthoseofBogfjellmoetal.suggestthat

inter-nal noise reducesto adult-like levelsby approximately

6–7years, whileage-relatedchangesinsamplingfollow

amoreextendedtrajectory.Ourfindingthatinternalnoise

reduceswithageechoesapreviousstudyintheauditory

domainwhichreportedhigherinternalnoiseinchildren

aged 6–11yearscompared toadults(Bussetal., 2006),

aswellasreportsofincreasedlevelsofinternalnoisein

infants(SkoczenskiandNorcia,1998)

Theequivalentnoisemethodgivesusanestimateofthe totalamountofinternalnoise,whilstremainingagnostic aboutitsprecisesource.However,wespeculatethathigh levelsofinternalnoiseinourdirectionintegrationtaskmay reflectimmaturityintheresponsesofdirection-sensitive cellsinV1.Specifically,imprecisioninestimatingthe direc-tionsoflocalelementsmaybeduetobroadbandwidthsof V1neuronsinchildrenbelowtheageof7years,whichlater narrowwithdevelopment(atleastintheprimatebrain,

Hatta etal.,1998).Conversely,developmental increases

insamplingmay reflectthe developmentofneuronsin higherareasofthemotionprocessinghierarchythought

tobeinvolvedinintegratinglocalmotionsignals,suchas MT/V5(BornandTootell,1992;Brittenetal.,1992).While

MTneuronsareresponsivetodirectioninformationand myelinatedatbirthinprimates(Flechsig,1901;Movshon

etal.,2004),theyshowimmaturitiesintheirintegrative properties(Movshonetal.,2004),whichcouldunderliethe extendeddevelopmentofsamplingreportedhere Further-more,thefactthatinternalnoisematuresbeforesampling corroboratesneurophysiologicalresearchshowingthatV1 maturesearlierthanextrastriateareas(Distleretal.,1996;

Gogtayetal.,2004;Houetal.,2009;Kourtzietal.,2006), whichhasbeenlinkedtodifferencesinsynapticpruning (Distleretal.,1996;Gogtayetal.,2004).Futurework com-biningpsychophysicalandneurophysiologicalmeasuresis necessaryto determinethepreciseneural substratefor theseeffects

Ourfindingsofage-relatedreductionsininternalnoise contrastsharplywithMcIntosh etal.’s (2008)reportof increasingneuralnoisemeasuredbyintra-participantEEG variabilitybetweentheagesof8and12years.McIntosh

etal.suggestedthatincreasingneuralnoisereflectedthe brain’s increasingcomplexity withage,allowingone to exploremultiplestatesandadapttodifferentsituations Thissortofcomplexity,however,isnotbeingtappedbythe visualintegrationtaskusedhere,andinstead,wereferto internalnoiseasuncertaintyinthecodingoflocalmotion directions Indeed, there are many different sources of noisewithinthenervoussystem(Faisaletal.,2008)andit

ispossiblethatnoisemayhavedifferenteffectsatdifferent levelsofthecorticalhierarchy.However,current compu-tationalandneuralmodelsofnoisearebasedonanimal andhumanadultbrains.Itthereforeremainsachallengeto determineexactlyhowthesemodelsshouldbeappliedto thedevelopingbrain.Thediscrepancybetweenourresults andthoseofMcIntoshetal.highlighttheimportanceof specifyingwhatismeantbynoiseandthelevelatwhich

itisthoughttohaveaneffectwhenconstructing develop-mentalmodels

Ourfindingsaddtoabodyofliteratureshowinga rel-ativelyprotracteddevelopmentofsensitivitytocoherent motion(Gunnetal.,2002;Hadadetal.,2011).Ourresults suggestthatmotioncoherencethresholdsreachadult-like levelsbyapproximately9years,whichisslightlyearlier thanpreviousaccountsthathavesuggestedthatmaturityis reachedby10–11yearsofage(Gunnetal.,2002),or12–14 yearsofage(Hadadetal.,2011).Discrepanciesintheageat whichadult-likelevelsarereachedarelikelytobedueto differencesinarangeofstimulusparameters(Narasimhan andGiaschi,2012).Ourstudyalsoallowedustotestthe

Ngày đăng: 02/11/2022, 08:53

TỪ KHÓA LIÊN QUAN

TÀI LIỆU CÙNG NGƯỜI DÙNG

TÀI LIỆU LIÊN QUAN